Researchers have introduced WeVisDoc, a novel two-stage data-centric framework designed to enhance the robustness of end-to-end document parsing across diverse layouts and acquisition conditions. To address the limitations of biased training corpora, the first stage of the framework expands semantic, structural, and appearance coverage by utilizing heterogeneous datasets combined with structure-preserving degradation synthesis. The second stage implements a diagnostic approach using a held-out probe to identify residual errors within specific visual-structural clusters. These diagnostics allow for targeted data construction and the strategic reallocation of the target-token budget to address specific model weaknesses. The resulting WeVisDoc-4B model demonstrates significant performance improvements, achieving an Overall score of 95.38 on OmniDocBench v1.6 and a mean Overall score of 75.54 across three PureDocBench tracks. Notably, the model ranks first among compared end-to-end parsers across all four tested settings. The transition from Stage I to Stage II specifically bolstered performance for both 2B and 4B model variants, with the 4B model showing a 4.03-point increase on the Real Degraded track of the PureDocBench benchmark, highlighting the efficacy of the error-driven data refinement strategy.